josef-pkt
commented
7 years ago an extra issue one-sided LR-test with a single hypothesis (e.g. a dispersion parameter) signed square root LR test example mentioned in Yang, Zhao, James W. Hardin, and Cheryl L. Addy. 2009. “A Score Test for Overdispersion in Poisson Regression Based on the Generalized Poisson-2 Model.” Journal of Statistical Planning and Inference 139 (4): 1514–21. doi:10.1016/j.jspi.2008.08.018. dispersion test Poisson versus GP-2 (generalized Poisson PR #3727 GP-P allows over and under dispersion, so Poisson is in the interior of the parameter space of GP-P (related one-sided at the boundary should be the same as one-sided signed LR-test for a single parameter that can be at the boundary, IIRC) in terms of application: score test in the Poisson versus GP2 case is more powerful and less conservative than Wald and LR in examples of Yang, Hardin and Addy. (They only check alpha=0.05, score_test doesn't look conservative in a quick MC at alpha=0.01 or alpha=0.001)
compare_lr, compare_jtest, ANOVA and related e.g. anova type tests base on deviance or pearson chisquare for GLM
I don't see a summary issue, but many individual issues
30 oldest issue for compare_xxx methods, still open
2673 test_lr, where we estimate both models
1358 RLM, implemented in PR #1341
GLM:
825 request for anova_glm
2688 more general model selection in GLM
2688 lr-type for "quasi-"mle i.e. dispersion only, reference Cameron and Trivedi
1626 reference to Stata function and documentation
3363 misspecified models, sum or linear combination of chisquare random variables
2641 PR to move compare_lr_test higher up to base classes
general discussion on missing features in GLM, especially also related to scale handling
related profile likelihood